Three-way multi-label classification: A review, a framework, and new challenges

The multi-label classification task is more challenging than the degenerated case of single-label classification due to diversified uncertainty. Uncertainty in multi-label classification not only involves label dependency but also the inconsistency of label correlation and imbalanced label association. While three-way decision methods excel in characterizing multifaceted uncertainty, developing a well-established three-way decision- based framework for multi-label classification remains challenging. Based on the historical developments of decision-theoretic rough sets and sequential three-way decision, this paper presents a systematic review of representative three-way multi-label classification models. By analyzing the contributions from existing studies, the multifaceted uncertainty of multi-label classification is classified into label uncertainty, correlation uncertainty, and structure uncertainty. To effectively deal with the structure uncertainty, some essential subproblems are identified, and a general three-way-based framework called the multi-label sequential decision-theoretic threeway decision (ML-SD3WD) model is presented. The ML-SD3WD model sequentially handles three-way topic generation, three-way label assignment, and three-way label enhancement by integrating decision-theoretic rough set with sequential three-way decision. Furthermore, some emerging directions in customizing the MLSD3WD model are delineated. Finally, limitations from both the label side and feature side are discussed and corresponding solutions are offered for uncertainty-driven solutions in practical applications. The results of this review will offer a road map for knowledge discovery in multi-label classification.